A Research On Data-driven GPS Trajectory Data Cleaning

With the wide spread of smart phones and other GPS devices,large-scale collections of GPS trajectory data are available.These data record the traveling trajectories of people and various vehicles which have plentiful information,and have given rise to location-based services and many other areas such as urban planning,intelligent transportation,and athlete behavior analysis.However,due to various factors such as GPS devices failure,sensor error,transmission error and storage error,most raw trajectory data contain a lot of noises which,if not properly handled,may lead to negative impacts on downstream researches and utilization over these trajectory data.Existing approaches for trajectory data cleaning mainly have the following three shortcomings: 1)utilizing extra map information as references;2)having strict requirements for the quality of the raw trajectory data,such as uniform and high sampling rates;3)only applicable to some particular scenarios.The aformentioned shortcomings lead to many limitations when applying these existing trajectory data cleaning approaches.This paper focuses on noise cleaning for GPS trajectory data,and explores methods of noise detection for trajectory data with different qualities while not utilizing any extra map information.In this paper,we propose a data-driven model for trajectory data cleaning.The model analyses the distribution of historical trajectory data to extract road information from them,and then uses the extracted information for noise detection.This model can overcome some application limitations of existing approaches to some extent.It does not have many requirements for the raw trajectory data(such as high sampling rates),and have robustness to nonuniform sampling rates and data sparsity,which make the model more generally applicable.The main contributions of this paper are summarized as follows:1.We propose a method for road skeleton extraction from historical trajectorydata.Due to the problems of noises and nonuniform distribution of the historicaltrajectory data,not every point in the historical trajectories can be chosen as theroad skeleton point.We construct a gird system on the historical trajectory dataand select representative cells in the grid system according to some conditions.Then through some smoothing and connecting processings,we can get smooth and uniformly distributed road skeleton.2.We propose a safe area construction method for noise detection.The methoduses the extracted road skeleton points as anchors,and construct safe areas for eachanchor according to the local distribution of the historical trajectory points aroundeach anchor,thus the safe area can separate noises from the correct points.Then,with the extracted road skeleton and the safe areas as references,we can detect and clean the noises in the trajectory data.3.We conduct extensive experiments to verify the effectiveness and efficiency ofour methods.We apply our methods on several real trajectory dataset,and designevaluation metrics to verify our methods both visually and statistically.We also compare our methods with some existing approaches and analyse the results.